It is essential for modern power system to forecast electricity price and load accurately. However, due to the strong correlation between the electricity price and the load, if the mutual influence is not taken into account, the accuracy of the forecast will be reduced. In order to improve the prediction accuracy of existing methods, price and load relationship are considered and a deep recurrent neural networks model is proposed for price and load forecasting, that is sparse autoencoder nonlinear autoregressive network with exogenous inputs comprising of feature engineering and forecasting. Firstly, an efficient sparse autoencoder is proposed to improve the effectiveness of feature extraction by improving the original method. Secondly, the nonlinear autoregressive network is used to forecast the load and price. The ISONE and PJM big datas of power market are simulated and verified. Compared with cascaded Elman networks, sparse autoencoder nonlinear autoregressive network reduces the mean absolute error by 16% in load forecasting and 7% in price forecasting.